Nested Entity Recognition via Part-of-Speech Aware Heterogeneous Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing nested entity recognition methods face challenges such as high time and computational complexity, poor inter-domain transferability, and limited generalization ability due to manual rule-based approaches, statistical machine learning vulnerabilities, and exhaustive candidate entity enumeration in deep learning methods.
Innovation Solution
A Nested Named Entity Recognition method based on part-of-speech awareness using heterogeneous graph representation learning, where part-of-speech knowledge is introduced to initialize text features, and a part-of-speech path-based empty random walk algorithm is employed to sample neighbor node information, improving recognition accuracy and efficiency through the Deep Graph Library framework.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual rule-based approaches are used for nested entity recognition, then recognition accuracy can be improved through expert knowledge, but time consumption and labor intensity increase significantly
Solution Approach 1:
The patent replaces manual rule-based approaches with a deep learning model that automatically learns recognition patterns from data. The neural network substitutes the mechanical process of manual rule creation and application, achieving both high accuracy through learned features and efficiency through automated processing without requiring expert intervention for each recognition task.
Solution Approach 2:
The patent transforms the recognition system from using fixed manual rules to using learned parameters from training data. By changing from static expert-defined rules to dynamic parameters learned through optimization, the system achieves better accuracy while reducing time consumption through automated feature extraction and pattern recognition.
2Productivity
If statistical machine learning methods are used for nested entity recognition, then processing speed can be improved, but generalization ability deteriorates due to vulnerability to text corpus distribution
Solution Approach 1:
The patent applies preliminary action by pre-training the deep learning model on large-scale data before deployment. This pre-training phase allows the model to learn robust patterns and representations that improve generalization ability, while the trained model then processes new text efficiently during inference, achieving both speed and adaptability.
Solution Approach 2:
The patent introduces dynamics by using a flexible neural network architecture that can adapt to different text distributions. Unlike rigid statistical methods, the deep learning model dynamically adjusts its predictions based on learned patterns, improving generalization ability while maintaining processing speed through efficient forward propagation.
3Reliability
If deep learning methods with exhaustive candidate entity enumeration are used for nested entity recognition, then recognition comprehensiveness can be improved, but computational complexity and time complexity increase significantly
Solution Approach 1:
The patent extracts only the most relevant features and candidate entities needed for recognition, rather than processing all possible candidates. By extracting key linguistic features and focusing on probable entity spans, the model achieves comprehensive recognition while significantly reducing computational complexity and training time.
Solution Approach 2:
The patent applies partial action by processing only the most promising candidate entities and features rather than exhaustively evaluating all possibilities. The model focuses computational resources on high-probability regions of the search space, achieving comprehensive recognition of actual entities while avoiding the excessive computational cost of enumerating all possible candidates.
Data Source
AI summary
Disclosed are a Nested Named Entity Recognition method based on part-of-speech awareness, system, device and storage medium therefor. The method uses a BiLSTM model to extract a feature of text word data in order to obtain a text word depth feature, and each text word of text to be recognized is initialized into a corresponding graph node, and a text heterogeneous graph of the text to be recognized is constructed according to a preset part-of-speech path, the text word data of the graph nodes is updated by an attention mechanism, and the features of all graph nodes of the text heterogeneous graph are extracted using the BiLSTM model, and a nested named entity recognition result is obtained after decoding and annotating. The present disclosure can recognize ordinary entities and nested entities accurately and effectively, and enhance the performance and advantages of the nested named entity recognition model.